WO2022024302A1

WO2022024302A1 - Machine learning data generation device, machine learning device, machine learning model generation method, and program

Info

Publication number: WO2022024302A1
Application number: PCT/JP2020/029255
Authority: WO
Inventors: 良平鈴木; 剛横矢; 諒増村; 浩貴太刀掛
Original assignee: 株式会社安川電機
Priority date: 2020-07-30
Filing date: 2020-07-30
Publication date: 2022-02-03
Also published as: JP7408815B2; JPWO2022024302A1; US20230134186A1

Abstract

The present invention identifies parameters representing the internal state of target equipment and uses interpolated values of these parameters to easily obtain appropriate machine learning data. A machine learning data generation device comprising: a real time series information acquisition unit which acquires and associates real time series information indicating the operating state of target equipment with prescribed labels; a physical simulation execution unit which generates a plurality of sets of virtual time series information on the basis of a plurality of parameters representing the internal state of the target equipment; a parameter identification unit which identifies parameters on the basis of the plurality of sets of virtual time series information and the real time series information, and associates the parameters with labels; a parameter generation unit which generates new parameters and labels on the basis of the identified parameters; a virtual time series information generation unit which performs a physical simulation using the new parameters, and generates virtual time series information corresponding to a new internal state; and a machine learning data generation unit which associates the new parameters and associated labels with the virtual time series information corresponding to the new internal state to generate new machine learning data.

Description

Machine learning data generator, machine learning device, machine learning model generation method and program

The present invention relates to a machine learning data generation device, a machine learning device, a machine learning model generation method, and a program.

In a machine that performs physical movements, in order to operate using machine learning, it is necessary to perform machine learning using machine learning data (vibration waveforms, etc.) that are in line with various realistic modes. However, in order to obtain sufficient learning data, it is necessary to actually operate the actual machine under various conditions to perform machine learning, which may require a great deal of labor and time.

The problem to be solved by the present invention is machine learning data appropriately labeled by specifying parameters representing the internal state of the target device by using physical simulation without actually operating the target device. It is an object of the present invention to provide a machine learning data generation device, a machine learning device, a machine learning model generation method, and a program for acquiring the above.

The machine learning data generation device according to one aspect of the present invention includes a real-time series information acquisition unit that acquires real-time series information indicating an operating state of a target device, which is a machine or an electric circuit, in association with a predetermined label. A physical simulation execution unit that executes a physical simulation that generates multiple virtual time-series information by sequentially calculating virtual states after a unit time based on each of multiple parameters that represent the internal state of the target device. , One or more of the plurality of parameters are specified based on the plurality of virtual time series information and the real time series information, and are specified by the parameter specifying unit and the parameter specifying unit associated with the label. A virtual time corresponding to a new internal state by executing the physical simulation using the new parameter and the parameter generator that generates the label corresponding to the new parameter, and the new parameter. Machine learning data that generates new machine learning data by associating the virtual time series information generator that generates series information with the new parameters and the labels that correspond to the virtual time series information that corresponds to the new internal state. It has a generator and a generator.

Further, the machine learning data generation device according to another aspect of the present invention further has a parameter storage unit in which the parameters that can be taken for each type of the label are stored in association with the label, and the parameter generation unit. Generates the new parameter based on the parameter stored in the parameter storage unit.

Further, in the machine learning data generation device according to another aspect of the present invention, the parameter generation unit is based on the relationship between the new parameter and the parameter group corresponding to each label stored in the parameter storage unit. , The label corresponding to the new parameter is determined.

Further, in the machine learning data generation device according to another aspect of the present invention, the label indicates whether the operating state of the target device is normal or abnormal, and the parameter generation unit is abnormal. Selectively generate the new parameter corresponding to the label indicating that.

Further, in the machine learning data generation device according to another aspect of the present invention, the new parameter is represented by a position on a parameter distribution diagram showing the distribution of the parameter to which the label is associated, and the parameter generation unit. Generates a parameter represented by a position where a predetermined number of the parameters exist within a predetermined range on the parameter distribution map as the new parameter.

Further, in the machine learning data generation device according to another aspect of the present invention, the virtual time series information generation unit has the new state based on the result of the physical simulation executed by using the new parameter. It has GAN (Generative Adversarial Networks) that generates the above-mentioned virtual time series information.

Further, the machine learning device according to another aspect of the present invention inputs the virtual time series information based on the machine learning data generation device according to any one of the above and the machine learning data, and displays the label. It has a learning unit for learning a neural network model, which is a neural network as an output.

Further, in the method for generating a machine learning model according to another aspect of the present invention, real-time series information indicating the operating state of a target device which is a machine or an electric circuit is acquired in association with a predetermined label. A physical simulation that generates a plurality of virtual time-series information is executed by sequentially calculating the virtual state after a unit time based on each of the acquisition step and a plurality of parameters representing the internal state of the target device. Based on the physical simulation execution step, the plurality of virtual time series information, and the real time series information, one or more of the plurality of parameters are specified, and the parameter specifying step associated with the label and the parameter. Based on the parameters specified by the specific step, the new parameter and the parameter generation step that generates the label corresponding to the new parameter, and the physical simulation using the new parameter are executed to perform a new internal state. The virtual time series information generation step for generating the virtual time series information corresponding to the new internal state and the virtual time series information corresponding to the new internal state are associated with the new parameter and the corresponding label to generate new machine learning data. It has a machine learning data generation step to be generated, and a learning step to train a neural network model, which is a neural network that inputs the virtual time series information and outputs the label based on the machine learning data.

Further, the program according to another aspect of the present invention includes a real-time series information acquisition step of acquiring real-time series information indicating an operating state of a target device which is a machine or an electric circuit in association with a predetermined label, and the above-mentioned. A physical simulation execution step that executes a physical simulation that generates multiple virtual time-series information by sequentially calculating virtual states after a unit time based on each of multiple parameters that represent the internal state of the target device. , One or more of the plurality of parameters are specified based on the plurality of virtual time series information and the real time series information, and are specified by the parameter specifying step and the parameter specifying step to be associated with the label. A virtual time corresponding to a new internal state by executing the physical simulation using the new parameter and the parameter generation step of generating the label corresponding to the new parameter and the new parameter. Machine learning data that generates new machine learning data by associating the virtual time series information generation step that generates series information with the virtual time series information corresponding to the new internal state and the label corresponding to the new parameter. A computer is made to execute a generation step and a learning step of learning a neural network model, which is a neural network that inputs the virtual time series information and outputs the label based on the machine learning data.

According to the present invention, appropriate machine learning data can be easily obtained by specifying a parameter representing the internal state of the target device and using the interpolated value.

It is a functional block diagram which shows the whole structure of the machine learning apparatus including the machine learning data generation apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the hardware composition of the machine learning data generation apparatus and the machine learning apparatus. It is a figure which shows the ball screw mechanism which is an example of a target device. It is a figure which shows an example of real time series information and virtual time series information. It is a figure which shows another example of the real time series information and the virtual time series information. It is a figure which shows another example of the real time series information and the virtual time series information. It is a figure which shows another example of the real time series information and the virtual time series information. It is a figure which shows an example of a parameter and a label stored in a parameter storage part. It is a figure which shows an example of a parameter distribution diagram. It is a figure explaining GAN. It is a figure which shows an example of the virtual time series information including noise. It is a figure which shows another example of the virtual time series information including noise. It is a figure which shows another example of the virtual time series information including noise. It is a figure which shows another example of the virtual time series information including noise. It is a flow chart of the machine learning data generation method and the machine learning method by the machine learning data generation apparatus and the machine learning apparatus which concerns on embodiment of this invention.

Hereinafter, a preferred embodiment (hereinafter referred to as an embodiment) for carrying out the present invention will be described with reference to each figure.

FIG. 1 is a functional block diagram showing the entire configuration of the machine learning device 100 including the machine learning data generation device 102 according to the first embodiment of the present invention. Here, the "machine learning data generator" refers to a device that generates machine learning data, which is teacher data used for learning in a machine learning model in which supervised learning is performed, and the "machine learning device" is used. Represents a device that performs machine learning model learning using machine learning data.

The machine learning device 100 and the machine learning data generation device 102 may be physically prepared as independent devices, but the present invention is not limited to this. The machine learning device 100 and the machine learning data generation device 102 may be incorporated as a part of another machine or device, and may be appropriately configured by using the physical configuration of the other machine or device as needed. It may be one. More specifically, the machine learning device 100 and the machine learning data generation device 102 may be implemented by software using a general computer.

Further, the programs for operating the computer as the machine learning device 100 and the machine learning data generation device 102 may be integrated or may be executed independently. Further, the program may be incorporated into other software as a module. Further, the machine learning device 100 and the machine learning data generation device 102 may be constructed on a so-called server computer, and only the functions thereof may be provided to a remote location via a public telecommunications line such as the Internet.

FIG. 2 is a diagram showing an example of the hardware configuration of the machine learning device 100 and the machine learning data generation device 102. FIG. 2 describes a general computer, which includes a CPU 202 (Central Processing Unit) as a processor, a RAM 204 (Random Access Memory) as a memory, an external storage device 206, a display device 208, an input device 210, and an I / O 212. (Inpur / Output) is connected by the data bus 214 so that electric signals can be exchanged with each other. The computer hardware configuration shown here is an example, and other configurations may be used.

The external storage device 206 is a device such as an HDD (Hard Disk Drive) or SSD (Solid State Drive) that can statically record information. The display device 208 is a CRT (Cathode Ray Tube), a so-called flat panel display, or the like, and displays an image. The input device 210 is one or a plurality of devices for inputting information by the user, such as a keyboard, a mouse, and a touch panel. The I / O 212 is one or more interfaces for a computer to exchange information with an external device. The I / O 212 may include various ports for wired connection and a controller for wireless connection.

The program for making the computer function as the machine learning device 100 and the machine learning data generation device 102 is stored in the external storage device 206, read out in the RAM 204 as needed, and executed by the CPU 202. That is, the RAM 204 stores a code for realizing various functions shown as a functional block in FIG. 1 by being executed by the CPU 202. Even if the program is recorded and provided on an information recording medium such as an optical disk, a magneto-optical disk, or a flash memory that can be read by a computer, the program is provided via an external information communication line such as the Internet via the I / O 212. May be done.

The machine learning data generation device 102 has, as its functional configuration, a real time series information acquisition unit 106, a virtual time series information generation unit 110 including a physical simulation execution unit 108, a parameter identification unit 112, and a parameter storage unit 114. The parameter generation unit 116 including the parameter generation unit 116 and the machine learning data generation unit 118 are included. Further, the machine learning device 100 includes a machine learning data generation device 102 and a learning unit 120. The machine learning data generation device 102 is prepared in line with the target device 104, which will be described later. Further, the machine learning device 100 trains the machine learning model used by the target device 104.

The real-time series information acquisition unit 106 acquires real-time series information indicating the operating state of the target device 104, which is a machine or an electric circuit, in association with a predetermined label. Here, the target device 104 referred to in the present specification is a device that performs work that exerts some physical action on the target object 316. As a specific example, a case where the target device 104 is a ball screw mechanism 300 that linearly moves the target object 316 will be described with reference to FIG. Since the ball screw mechanism 300 itself is known, the explanation is kept to a minimum.

As shown in FIG. 3, the ball screw mechanism 300 includes a sensor 302, a control unit 304, a servo amplifier 306, a servo motor 308, a ball screw shaft 310, a ball screw nut 312, a table 314, and the like.

The sensor 302 acquires real-time series information indicating the operating state of each part of the ball screw mechanism 300. Specifically, for example, the sensor 302 is a torque sensor that detects the torque of the servomotor 308. Further, the sensor 302 is an angle sensor that acquires the rotation angle of the ball screw shaft 310 or the servomotor 308 as real-time series information, the position sensor that acquires the position of the table 314 as real-time series information, and the control unit 304 is the servomotor 308. It may be a voltage sensor or a current sensor that acquires the voltage and current output to the sensor as real-time series information.

Specifically, for example, the sensor 302 provides real-time series information on changes in the rotational torque of the servomotor 308 shown in FIGS. 4 (a), 5 (a), 6 (a), and 7 (a). Get as. Each figure is a diagram showing the rotational torque of the servomotor 308 when the operation of moving the table 314 twice in one direction and then moving twice in the opposite direction is repeatedly executed at 1-second intervals.

In this example, the ball screw mechanism 300 is in a normal state when the rotational torque converges substantially flat after each peak. That is, FIG. 4A is real time series information (hereinafter referred to as normal data) when the ball screw mechanism 300, which is the target device 104, is in a normal state. On the other hand, when the rotational torque does not converge flatly after each peak, the ball screw mechanism 300 is in an abnormal state. That is, FIGS. 5 (a), 6 (a), and 7 (a) show real-time series information when the ball screw mechanism 300, which is the target device 104, is in an abnormal state (hereinafter, abnormality data 1, respectively). 2 and 3).

The control unit 304 is a computer that controls the servo amplifier 306 that outputs current, voltage, etc. to the servo motor 308. The control unit 304 may acquire the output of the sensor 302 and perform feedback control with respect to the control to the servo amplifier 306. The servomotor 308 rotates the ball screw shaft 310 based on the current and voltage acquired from the servo amplifier 306 under the control of the control unit 304. The ball screw nut 312 is fitted to the ball screw shaft 310, and the ball screw shaft 310 rotates to move in the axial direction of the ball screw shaft 310. On the table 314, the object to be moved 316 is arranged and fixed to the ball screw nut 312 to move the object 316 in the axial direction of the ball screw shaft 310.

The real time series information acquisition unit 106 acquires real time series information representing the operating state of the ball screw mechanism 300 from the sensor 302 included in the ball screw mechanism 300. For example, the real-time series information acquisition unit 106 acquires the rotation angle of the ball screw shaft 310 or the servomotor 308 as real-time series information. Further, the real time series information acquisition unit 106 acquires a label in association with the real time series information. The label is, for example, a label indicating that the operating state represented by the associated real-time series information is normal or abnormal when the operating state represents the normal or abnormal operation of the target device 104.

Specifically, for example, the real-time series information acquisition unit 106 obtains the real-time series information shown in FIGS. 4 (a), 5 (a), 6 (a), and 7 (a) from the target device 104. get. Further, the real-time series information acquisition unit 106 acquires a label indicating the state of the ball screw mechanism 300 when the real-time series information is acquired in association with the real-time series information. In the above example, the real-time series information acquisition unit 106 acquires a label indicating that it is normal in association with the real-time series information shown in FIG. 4 (a). Further, the real-time series information acquisition unit 106 acquires a label indicating that the information is abnormal in association with the real-time series information shown in FIGS. 5 (a), 6 (a), and 7 (a).

The label associated with the acquired real-time series information may be determined by the judgment of the user who has seen the real-time series information, or by the judgment of the user according to the inspection result of a given inspection device. It may be decided without. Further, the label may be a label indicating excellent, good, acceptable or unacceptable when the operating state represented by the associated real time series information is evaluated stepwise.

The target device 104 is not limited to the ball screw mechanism 300 as long as it is a device that acquires real-time series information indicating the operating state of the target device 104. For example, the target device 104 includes parts and parts pickup, parts mounting (for example, fitting a bearing into a housing, screw fastening, etc.), packing (boxing of foods such as confectionery, etc.), and various processing (deburring, etc.). It may be an automatic machine that repeatedly and continuously performs various operations such as metal processing such as polishing, molding and cutting of soft materials such as food, resin molding and laser processing), painting and cleaning.

The virtual time series information generation unit 110 includes the physics simulation execution unit 108. The physics simulation execution unit 108 generates a plurality of virtual time-series information by sequentially calculating virtual states after a unit time based on each of a plurality of parameters representing the internal states of the target device 104. To execute. The virtual time-series information generation unit 110 executes a physics simulation by the physics simulation execution unit 108, and generates virtual time-series information corresponding to the internal state.

The physics simulation execution unit 108 according to the present embodiment is constructed as specifically conforming to the target device 104 that performs a specific physical operation. The content of the physical operation and the use of the target device 104 are not particularly limited. For example, the physics simulation execution unit 108 according to the present embodiment executes the physics simulation based on the model of the ball screw mechanism 300.

Specifically, as a virtual model of the actual ball screw mechanism 300 in the virtual space of the physical simulation execution unit 108, the sensor 302, the control unit 304, the servo amplifier 306, the servo motor 308, the ball screw shaft 310, and the ball screw nut A virtual model of the ball screw mechanism 300 constructed in each part of 312 and the table 314 is prepared in advance. The virtual model of the ball screw mechanism 300 includes the motor rotation angle (θ _m ), the position of the ball screw nut 312 in the ball screw axis 310 direction at time t (x _t ), the rotation torque (T _m ) of the servo motor 308, and the rotation. The inertial moment (J _r ) of the system, the total mass (M _t ) of the weights of the driven body, that is, the ball screw nut 312, the table 314 and the object 316 placed on the table 314, and the viscous friction coefficient of the rotating system (M t). Parameters such as D _r ), viscous friction coefficient (C _t ) of linear motion guide, diameter (R) of ball screw shaft 310, and axial rigidity (K _t ) of drive mechanism are included.

The physics simulation execution unit 108 simulates the physical operation performed by the target device 104 in the virtual space by operating the virtual model including the above parameters according to the virtual operation command. The behavior of each part of the ball screw mechanism 300 in the virtual space naturally reproduces the situation when the actual ball screw mechanism 300 is operated.

The physics simulation execution unit 108 is constructed to output each parameter after a unit time using the above parameters at a certain point in time. Specifically, the parameters representing the internal state include parameters that change with time. For example, when the servomotor 308 applies torque to the ball screw shaft 310, the ball screw shaft 310 rotates and the position of the ball screw nut 312 in the direction of the ball screw shaft 310 changes.

The physics simulation execution unit 108 outputs each parameter after a unit time using the above parameters at a certain point in time. Then, the physics simulation execution unit 108 uses each parameter after the unit time and further outputs each parameter after the unit time. In this way, the physics simulation execution unit 108 performs the calculation using each parameter after the unit time, and then using the parameter that changes sequentially by using the parameter in the calculation. Then, the virtual time-series information generation unit 110 generates virtual time-series information corresponding to the internal state of the target device 104 by integrating each parameter for each unit time output by the physics simulation execution unit 108 on the time axis. do.

An example of virtual time-series information generated by the virtual time-series information generation unit 110 will be described. Here, a predetermined initial value is set as the above parameter, and the virtual model of the ball screw mechanism 300 moves the table 314 twice in one direction and then moves twice in the opposite direction at 1-second intervals. The case where the operation is performed according to the virtual operation command to be performed will be described.

4 (b), 5 (b), 6 (b) and 7 (b) show the viscous friction coefficient D _r of the rotary system and the viscous friction coefficient C _t of the linear motion guide given as initial parameters. It is virtual time series information which represents the time-dependent change of the torque of the generated servomotor 308 under different conditions.

FIG. 4B shows virtual time-series information of the torque of the servomotor 308 generated when the viscous friction coefficient D _r of the rotary system is 0.003 and the viscous friction coefficient C _t of the linear motion guide is 50. be. FIG. 5B is virtual time-series information of the torque of the servomotor 308 generated when the viscous friction coefficient D _r of the rotary system is 0.0035 and the viscous friction coefficient C _t of the linear motion guide is 1000000. be. FIG. 6B is virtual time-series information of the torque of the servomotor 308 generated when the viscous friction coefficient D _r of the rotary system is 0.3 and the viscous friction coefficient C _t of the linear motion guide is 50. be. FIG. 7B is virtual time-series information of the torque of the servomotor 308 generated when the viscous friction coefficient D _r of the rotary system is 0.1 and the viscous friction coefficient C _t of the linear motion guide is 1000. be.

The physics engine used for the physics simulation may be one that corresponds to the assumed physical work. When the operation of the ball screw mechanism 300 is assumed as in this example, a physics engine capable of performing collision determination and dynamic simulation may be selected or constructed. Different physics will, of course, select or build a physics engine that simulates fluid simulations, fracture simulations, and all other physics. For example, when the target device 104 is an electronic circuit, the physics simulation execution unit 108 outputs an electric signal represented by a voltage or current that is sequentially changed from the electronic circuit based on a command input to the electronic circuit. , May be generated as virtual time series information.

The parameter specifying unit 112 identifies one or more parameters from a plurality of parameters based on a plurality of virtual time series information and real time series information indicating an operating state, and associates them with a label. Specifically, for example, first, the virtual time-series information generation unit 110 randomly creates a value within the range of values that each parameter can physically take, and generates virtual time-series information based on the created parameters. The parameter specifying unit 112 has the smallest error with the real time series information acquired from the target device 104 by the real time series information acquisition unit 106 from among the plurality of virtual time series information generated by the virtual time series information generation unit 110. Identify virtual time series information. As a result, the parameter specifying unit 112 specifies the parameter corresponding to the specified virtual time series information. Further, the parameter specifying unit 112 associates the specified virtual time series information with the same label as the label associated with the real time series information acquired from the target device 104.

For example, the parameter specifying unit 112 has the real time series information of the torque of the servo motor 308 shown in FIG. 4 (a) and the virtual time series information of the smallest error of the torque of the servo motor 308 shown in FIG. 4 (b). Identify virtual time series information. Then, the parameter specifying unit 112 is associated with the real time series information of the torque of the servo motor 308 shown in FIG. 4 (a) and the virtual time series information of the torque of the servo motor 308 shown in FIG. 4 (b). Associate the label "normal".

Similarly, in the parameter specifying unit 112, the real time series information shown in FIGS. 5 (a), 6 (a), and 7 (a) and the virtual time series information having the smallest error are sequentially shown in FIG. 5 (b). ), The virtual time series information shown in FIGS. 6 (b) and 7 (b) is specified. Further, the parameter specifying unit 112 refers to the virtual time series information shown in FIGS. 5 (b), 6 (b), and 7 (b) with respect to FIGS. 5 (a), 6 (a), and 7 (b). Associate the label "abnormal" associated with (a).

If one parameter cannot be specified from a plurality of parameters, the parameter specifying unit 112 may specify two or more parameters. For example, the plurality of virtual time-series information generated by the virtual time-series information generation unit 110 may include a plurality of virtual time-series information having an error as small as that of the real time-series information. In this case, the parameter specifying unit 112 specifies each parameter corresponding to the plurality of virtual time series information. Further, the parameter specifying unit 112 associates the plurality of virtual time series information with the same label as the label associated with the real time series information acquired from the target device 104.

The parameter storage unit 114 stores the specified parameter in association with the label associated with the parameter. Specifically, the parameter storage unit 114 associates the real time series information acquired by the real time series information acquisition unit 106 with the parameters specified as the parameters corresponding to each real time series information and each real time series information. The attached label and the associated label are stored.

For example, as shown in FIG. 8, the parameter storage unit 114 has the real time series information shown in FIGS. 4 (a), 5 (a), 6 (a), and 7 (a), and the real time series information. The viscous friction coefficient D _r of the rotary system and the viscous friction coefficient C _t of the linear motion guide used when generating each virtual time series information specified as the virtual time series information with the smallest error, and each real time series. Remember the "normal" or "abnormal" label associated with the information. The values 1 to 4 in the real-time series information field in FIG. 8 correspond to the real-time series information shown in FIGS. 4 (a), 5 (a), 6 (a), and 7 (a), respectively. do.

Further, the parameter storage unit 114 stores parameters that can be taken for each type of label in association with the label. Specifically, FIG. 9 is a diagram showing the relationship between the viscous friction coefficient D _r of the rotary system stored in the parameter storage unit 114, the viscous friction coefficient C _t of the linear motion guide, and the label. Note that each real time-series information shown in FIG. 9 is all real-time-series information acquired by the real-time-series information acquisition unit 106, and includes virtual time-series information generated by the virtual time-series information generation unit 110. No. As shown in FIG. 9, in a two-dimensional plane in which the vertical axis is the viscous friction coefficient D _r of the rotary system and the horizontal axis is the viscous friction coefficient C _t of linear motion guidance, the real-time series information in which different labels are associated is , Is distributed unevenly in a predetermined area. Hereinafter, a diagram showing the distribution of parameters associated with each label as shown in FIG. 9 is referred to as a parameter distribution map.

That is, the real-time series information (normal data) associated with the "normal" label is distributed inside the elliptical region of FIG. 9, and the real-time series information (abnormal data 1, abnormal data 1) associated with the "abnormal" label is distributed. 2 and 3) are distributed outside the elliptical region of FIG. The distribution is such that the viscous friction coefficient D _r of the rotary system of the ball screw mechanism 300 and the viscous friction of the linear motion guide are determined according to whether the ball screw mechanism 300 is operating normally or in an abnormal state. This is due to the difference in the relationship with the coefficient C _t .

Therefore, the parameters that can be associated with the normal label are the parameters that are distributed inside the elliptical region of FIG. 9, and the parameters that can be associated with the abnormal label are the parameters that are distributed outside the elliptical region of FIG. Is. In this way, the parameter storage unit 114 stores the parameters that can be taken for each type of label in association with the label.

Note that, in FIG. 9, the boundaries of the real time series information associated with different labels are shown by ellipses, but the boundaries may be set by any method. For example, the position where the sum of the square of the distance to the closest normal data and the square of the distance to the closest abnormal data is the smallest may be set as the boundary. Further, the position of the boundary may be set according to the number of normal data or abnormal data existing within a predetermined distance. Also, the boundaries do not have to be clearly determined.

The parameter generation unit 116 generates a new parameter and a label corresponding to the new parameter based on the parameter specified by the parameter specifying unit 112. For example, the parameter generation unit 116 generates new parameters based on the parameters stored in the parameter storage unit 114 as shown in FIGS. 8 and 9.

Specifically, the parameter generation unit 116 generates the viscous friction coefficient D _r of the rotational system corresponding to the position where the data does not exist and the viscous friction coefficient C _t of the linear motion guide as new parameters in the parameter distribution map. .. The parameter is a parameter representing a new internal state that is not reproduced in the actual ball screw mechanism 300.

Further, the parameter generation unit 116 determines the label corresponding to the new parameter according to the relationship between the new parameter and the parameter group corresponding to each label stored in the parameter storage unit 114. Specifically, for example, the parameter group existing inside the ellipse of FIG. 9 corresponds to the "normal" label, and the parameter group existing outside the ellipse of FIG. 9 corresponds to the "abnormal" label. Therefore, the parameter generation unit 116 determines that the label corresponding to the new parameter is "normal" when the generated new parameter is inside the ellipse of FIG. Further, the parameter generation unit 116 determines that the label corresponding to the new parameter is "abnormal" when the generated new parameter exists outside the ellipse of FIG.

Note that, depending on the acquired real time-series information, the areas in which each label stored in the parameter storage unit 114 and the corresponding parameter group are distributed may overlap with each other. That is, the boundary line shown in FIG. 9 may not be clearly determined. In this case, the parameter generation unit 116 evaluates the relationship between the new parameter and each label stored in the parameter storage unit 114 and the corresponding parameter group, and based on the evaluation result, the new parameter and the corresponding label are generated. You may decide. For example, the parameter generation unit 116 may calculate a probability distribution function with a position in the parameter distribution map as a variable for each label. Then, the parameter generation unit 116 may determine the label having the highest probability of being associated with the newly generated parameter as the label corresponding to the new parameter, based on the probability distribution function. As a result, even when the boundary line shown in FIG. 9 cannot be clearly determined, an appropriate label can be determined as a label corresponding to the new parameter.

Further, the parameter generation unit 116 may evaluate the number of parameters existing within a predetermined distance centered on the new parameter for each label. In the example of the parameter distribution diagram shown in FIG. 9, the parameter generation unit 116 evaluates the number of parameters associated with the normal label and the abnormal label existing within a predetermined distance centered on the new parameter. Then, the label associated with the large number of parameters may be determined as the label corresponding to the new parameter.

Further, the parameter generation unit 116 may selectively generate a new parameter corresponding to the label indicating that it is abnormal. Specifically, for example, the parameter generation unit 116 may selectively generate new parameters so that the positions of the new parameters on the parameter distribution map shown in FIG. 9 are located outside the ellipse. In the parameter distribution map, the parameters located outside the ellipse are associated with the label representing "abnormality". Therefore, the new parameters generated are associated with a label indicating anomalies. Normally, it is often more difficult to collect abnormal data than normal data because the state in which the target device operates abnormally is not reproducible. Abnormality data can be efficiently collected by selectively generating a new parameter corresponding to the label indicating that the parameter generation unit 116 is abnormal.

Further, the parameter generation unit 116 may generate a parameter represented by a position where a predetermined number of parameters exist within a predetermined range on the parameter distribution map as a new parameter. Specifically, as shown in FIG. 9, the new parameter is represented by a position on the parameter distribution map that represents the distribution of the parameter to which the label is associated. The parameter generation unit 116 determines the position where a predetermined number of parameters exist in the surroundings as the position of the newly generated parameter. For example, the viscous friction coefficient (D _r ) of the rotating system is in the range of ± 0.1 and the viscous friction coefficient (C _t ) of the linear motion guide is in the range of ± 1000 around the position of the new parameter. The parameter generation unit determines the position of the newly generated parameter so that there are 10 or more parameters. This makes it possible to prevent new parameters from being generated at positions extremely distant from the positions of the parameters specified based on the real time series information. Therefore, new parameters that are more realistic are generated. The above range and number are examples. New parameters may be generated such that a predetermined number of parameters exist within a predetermined distance centered on the position of the new parameter.

The machine learning data generation unit 118 generates new machine learning data by associating the virtual time series information corresponding to the new internal state with the new parameter and the corresponding label. Specifically, the machine learning data generation unit 118 is determined by the virtual time series information generated in the virtual time series information generation unit 110 based on the new parameters generated by the parameter generation unit 116 and the parameter generation unit 116. New machine learning data is generated by associating the relevant parameter with the corresponding label.

The machine learning data generation unit 118 may include the GAN1000. GAN1000 (Generative Adversarial Networks) generates virtual time series information that is indistinguishable from real time series information. Here, the GAN1000 will be briefly described with reference to FIG. Since the GAN1000 is known, the explanation is kept to a minimum.

The GAN1000 has the configuration shown in FIG. 10 and includes two neural networks called a generator 1002 and a discriminator 1004. The generator 1002 receives the input of the virtual time series information and the predetermined noise, and outputs the virtual time series information including the noise. On the other hand, both the virtual time-series information including noise generated by the generator 1002 and the real-time-series information acquired from the target device 104 are input to the discriminator 1004. At this time, the discriminator 1004 is not informed whether the input data is virtual time series information or real time series information.

The output of the discriminator 1004 determines whether the input data is virtual time-series information or real-time-series information. Then, the GAN1000 correctly discriminates between some virtual time-series information and real-time-series information prepared in advance by the discriminator 1004, and in the generator 1002, both of them are used in the discriminator 1004. Reinforcement learning is repeated so that it cannot be discriminated.

As a result, the discriminator 1004 will not be able to distinguish between the two (for example, if the same number of virtual time series information and real time series information are prepared, the correct answer rate will be 50%). In this state, the generator 1002 outputs the virtual time-series information as if it were the actual real-time-series information, which is indistinguishable from the real-time-series information, based on the virtual time-series information. Therefore, the machine learning data generation unit 118 generates machine learning data based on the virtual time series information including noise generated by using the generator 1002 and the discriminator 1004 for which the above learning has been executed. You may.

Specifically, for example, the virtual time series information shown in FIGS. 11 (a), 12 (a), 13 (a), and 14 (a) is input to the GAN 1000 included in the machine learning data generation unit 118. In this case, the virtual time series information including the noise shown in FIGS. 11 (b), 12 (b), 13 (b), and 14 (b) is output.

Here, FIGS. 11 (a), 12 (a), 13 (a), and 14 (a) are shown in FIGS. 4 (b), 5 (b), 6 (b), and 7 (a), respectively. This is the virtual time series information shown in b). Since the virtual time-series information generated by the virtual time-series information generation unit 110 does not include noise, it has a shape that is difficult to acquire from the actual target device 104. However, it is difficult at first glance to distinguish the virtual time-series information including noise generated by the GAN1000 from the real-time-series information. That is, according to GAN1000, it is possible to generate more realistic virtual time series information.

As described above, the machine learning data generator 102 makes it easy to obtain a large number of different machine learning data within a practical time and cost range. Even if the failure that occurs in the target device is infrequent, when the internal state of the target device in which the failure of the mode has occurred is represented by executing the physical simulation using the parameters that reflect the mode of the failure. Series information can be used as training data.

The machine learning device 100 includes the above-mentioned machine learning data generation device 102 and a learning unit 120. The learning unit 120 trains a neural network model, which is a neural network that inputs virtual time series information and outputs labels based on machine learning data. Specifically, n machine learning data generated by the machine learning data generation unit 118 are input to the neural network. Here, n is a sufficient number for machine learning and is appropriately set. Further, it is assumed that i is an integer from 1 to n, and the machine learning data i includes virtual time series information i and a label i. In the neural network, virtual time series information i is input and a score is calculated.

Here, the score is a value indicating the degree of agreement with a predetermined label (“normal” or “abnormal” label in the above example), and is, for example, an output value of CNN. Then, the comparison result (hereinafter, error) between the label i input to the learning unit 120 in association with the virtual time series information i and the score is specified. The error may be data having a value of 0 or more and 1 or less. The error may be, for example, data that takes 1 as a value when the calculated score and the label i match, and 0 as a value when they do not match.

Further, based on the error, the weighting factor between each node of the CNN is updated by, for example, the error back propagation method. The neural network changes i from 1 to n and repeatedly updates the weighting coefficient each time machine learning data is input. As a result, the learning of the learning unit 120 is executed.

FIG. 15 is a flow chart of a machine learning data generation method and a machine learning method by the machine learning device 100 and the machine learning data generation device 102 according to the present embodiment. Among the flows shown in the figure, S1502 to S1514 correspond to the machine learning data generation method, and S1502 to S1516 correspond to the machine learning method.

First, the real-time series information acquisition unit 106 acquires real-time series information indicating the operating state of the target device 104, which is a machine or an electric circuit, in association with a predetermined label (S1502). Next, the physics simulation execution unit 108 generates a plurality of virtual time-series information by sequentially calculating the virtual state after a unit time based on each of the plurality of parameters representing the internal state of the target device 104. Perform a physics simulation. As a result, the virtual time-series information generation unit 110 generates virtual time-series information corresponding to the internal state (S1504).

Next, the parameter specifying unit 112 identifies one or more parameters from the plurality of parameters based on the plurality of virtual time series information generated in S1504 and the real time series information indicating the operating state, and associates them with the label. (S1506). The identified parameter and the label associated with the parameter are stored in the parameter storage unit 114. Further, at this time, by calculating the boundary of each label on the parameter distribution map and the probability distribution function, the parameters that can be taken for each type of label may be stored in the parameter storage unit 114 in association with the label. ..

Next, the parameter generation unit 116 generates a new parameter and a label corresponding to the new parameter based on the parameter specified by the parameter specifying unit 112 (S1508). Then, the virtual time-series information generation unit 110 executes a physics simulation using the new parameters generated in S1508, and generates virtual time-series information corresponding to the new internal state (S1510). The machine learning data generation unit 118 associates a label with the virtual time series information corresponding to the new internal state generated in S1510, and generates new machine learning data (S1512). The generated machine learning data is sequentially accumulated.

In S1514, it is determined whether or not the number of accumulated machine learning data is sufficient. If the number of machine learning data is not sufficient (S1514: N), the process returns to S1508 and the machine learning data is repeatedly generated. If the number of records is sufficient (S1514: Y), the process proceeds to S1518. The target number of required machine learning data may be set in advance. Alternatively, the result of machine learning in S1516 may be evaluated, and if the learning is not sufficient, S1508 to S1514 may be executed again to additionally generate machine learning data. The evaluation of the machine learning result may be performed by evaluating the convergence of the internal state of the neural network model in the learning unit 120, or by inputting test data into the neural network model and using the correct answer rate of the obtained output. You may.

Next, the learning unit 120 executes learning of the neural network model according to the achievement status based on the generated machine learning data. In this way, in this embodiment, a trained neural network model is obtained.

As is clear from the example of the ball screw mechanism 300 shown in this example, it is not easy to actually prepare a sufficient number of appropriate machine learning data for sufficiently training the neural network model of the learning unit 120. .. The ball screw mechanism 300 is configured by combining a plurality of parts, and when the parts causing the failure are different, or when the parts causing the failure are the same but the cause of the failure is different, various different failure modes are used. Is to appear. In order to acquire machine learning data from the ball screw mechanism 300 operating in the various different failure modes, it is necessary to realize the various different failure modes by using the actual ball screw mechanism 300, but it is too large. It is not realistic because it takes time and cost.

However, according to the present embodiment, machine learning is performed without actually requiring preparation for the target device 104 having a different failure mode. Therefore, the machine learning data generation device 102 according to the present embodiment virtually executes the physical operation by the target device 104 to generate a sufficient number of machine learning data for the neural network model in a realistic time and cost. Can be generated. Further, the machine learning device 100 according to the present embodiment can train the neural network model from the generated machine learning data.

In addition, there is a technology called VAE that captures the characteristics of machine learning data, encodes it into latent variables, and generates data similar to machine learning data again from the latent variables. According to VAE, the target device 104 can generate machine learning data without actually performing physical operation. For example, the VAE converts machine learning data such as the output of the sensor 302 provided in the machine into a feature vector, and generates new teacher data by interpolating these. However, VAE only interpolates data that represents the superficial state of the machine, so if the internal state of the machine is only slightly different but the data output from the machine is significantly different, a suitable label. Cannot generate machine learning data with.

However, according to the present embodiment, since the physics simulation execution unit 108 executes the physics simulation based on a plurality of parameters representing the internal states of the target device 104, machine learning that reflects the operation performed by the target device 104 is performed. Can be done. Further, since the parameter represents the internal state of the target device 104, the range of values that the parameter can take is physically determined. Therefore, it is possible to prevent learning from being executed based on virtual time-series information generated from parameters that are not realistically possible.

100 machine learning device, 102 machine learning data generation device, 104 target device, 106 real time series information acquisition unit, 108 physical simulation execution unit, 110 virtual time series information generation unit, 112 parameter identification unit, 114 parameter storage unit, 116 parameters Generation unit, 118 machine learning data generation unit, 120 learning unit, 202 CPU, 204 RAM, 206 external storage device, 208 display device, 210 input device, 212 I / O, 214 data bus, 300 ball screw mechanism, 302 sensor, 304 control unit, 306 servo amplifier, 308 servo motor, 310 ball screw shaft, 312 ball screw nut, 314 table, 316 object, 1000 GAN, 1002 generator, 1004 discriminator.

Claims

A real-time-series information acquisition unit that acquires real-time-series information indicating the operating state of a target device, which is a machine or an electric circuit, in association with a predetermined label.
A physics simulation execution unit that executes a physics simulation that generates a plurality of virtual time-series information by sequentially calculating virtual states after a unit time based on each of a plurality of parameters representing the internal states of the target device. When,
Based on the plurality of virtual time-series information and the real-time-series information, a parameter specifying unit that identifies one or more of the plurality of parameters and associates them with the label.
A parameter generation unit that generates a new parameter and the label corresponding to the new parameter based on the parameter specified by the parameter identification unit.
A virtual time-series information generator that executes the physics simulation using the new parameters and generates virtual time-series information corresponding to the new internal state.
A machine learning data generation unit that generates new machine learning data by associating the new parameter with the label corresponding to the virtual time series information corresponding to the new internal state.
Machine learning data generator with.
Further, the parameter that can be taken for each type of the label has a parameter storage unit that is stored in association with the label.
The parameter generation unit generates the new parameter based on the parameter stored in the parameter storage unit.
The machine learning data generation device according to claim 1.
The parameter generation unit determines the label corresponding to the new parameter by the relationship between the new parameter and the parameter group corresponding to each label stored in the parameter storage unit according to claim 2. The machine learning data generator described.
The label indicates whether the operating state of the target device is normal or abnormal.
The parameter generator selectively generates the new parameter corresponding to the label indicating that it is abnormal.
The machine learning data generation device according to claim 2.
The new parameter is represented by a position on the parameter distribution map that represents the distribution of the parameter with which the label is associated.
The parameter generation unit generates a parameter represented by a position where a predetermined number of the parameters exist within a predetermined range on the parameter distribution map as the new parameter.
The machine learning data generation device according to claim 3.
The virtual time-series information generation unit has GAN (Generative Adversarial Networks) that generates the virtual time-series information in the new state based on the result of the physical simulation executed using the new parameters. The machine learning data generation device according to any one of claims 1 to 5.
The machine learning data generator according to any one of claims 1 to 6.
A learning unit that trains a neural network model, which is a neural network that inputs the virtual time series information and outputs the label based on the machine learning data.
Machine learning device with.
A real-time-series information acquisition step for acquiring real-time-series information indicating the operating state of a target device, which is a machine or an electric circuit, in association with a predetermined label.
A physics simulation execution step of executing a physics simulation that generates a plurality of virtual time-series information by sequentially calculating a virtual state after a unit time based on each of a plurality of parameters representing the internal state of the target device. When,
A parameter specifying step of identifying one or more of the plurality of parameters based on the plurality of virtual time-series information and the real-time series information and associating the label with the label.
A parameter generation step that generates a new parameter and the label corresponding to the new parameter based on the parameter specified by the parameter specifying step.
A virtual time-series information generation step that executes the physics simulation using the new parameters and generates virtual time-series information corresponding to the new internal state.
A machine learning data generation step for generating new machine learning data by associating the new parameter with the label corresponding to the virtual time series information corresponding to the new internal state.
Based on the machine learning data, a learning step for training a neural network model, which is a neural network in which the virtual time series information is input and the label is output,
How to generate a machine learning model with.
A real-time-series information acquisition step for acquiring real-time-series information indicating the operating state of a target device, which is a machine or an electric circuit, in association with a predetermined label.
A physics simulation execution step of executing a physics simulation that generates a plurality of virtual time-series information by sequentially calculating a virtual state after a unit time based on each of a plurality of parameters representing the internal state of the target device. When,
A parameter specifying step of identifying one or more of the plurality of parameters based on the plurality of virtual time-series information and the real-time series information and associating the label with the label.
A parameter generation step that generates a new parameter and the label corresponding to the new parameter based on the parameter specified by the parameter specifying step.
A virtual time-series information generation step that executes the physics simulation using the new parameters and generates virtual time-series information corresponding to the new internal state.
A machine learning data generation step for generating new machine learning data by associating the new parameter with the label corresponding to the virtual time series information corresponding to the new internal state.
Based on the machine learning data, a learning step for training a neural network model, which is a neural network in which the virtual time series information is input and the label is output,
A program that lets your computer run.